Agent onboarding exam: Fable 5's most difficult task still requires a blank paper, with single question costs 4 to 12 times higher According to Beating monitoring, the University of California, Berkeley's RDI has led and collaborated with hundreds of industry experts to launch a new AI agent evaluation benchmark, Agents' Last Exam (ALE), to assess the ability of agents to complete real digital professional work. ALE covers 55 sub fields of digital expertise and collects over 1500 validation tasks from actual projects of human experts, supporting result validation in GUI and CLI interactive environments. The first batch of tests covered cutting-edge systems such as Fable 5, GPT-5.5, and Composer 2.5. According to the latest official website comparison, in the most difficult tasks that require continuous reasoning and deep professional knowledge, the success rate of all tested agents is 0%. Fable 5, which was just released this week, also failed the test. This is mainly because the review triggered security policies, and about 35% of Fable 5's tasks were rolled back and switched to the old version Opus 4.8, resulting in overall performance that was far less eye-catching than other rankings. In terms of single task API cost, Fable 5 is about $15.70, far higher than GPT-5.5's $3.80 and Composer 2.5's $1.33, with expenses 4 to 12 times higher for the same task. The test also found that the most common reason for the failure of intelligent agents is to declare success too early, hastily ending work without actual verification results, or even missing files or miscalculated data. The evaluation team has simultaneously released a subset ALE-CLI for command-line agents. Compared with existing Terminal Bench and SWE bench Pro, ALE-CLI covers 40 sub domains, and the average human time for a single task can reach several hours or even weeks. In the command-line evaluation, the best performing intelligent agent had a pass rate of only 25.2%. The evaluation team pointed out that the era of useful intelligent agents has arrived, but there is still a long way to go before they can truly replace humans. [Original link]